A Method, Device, and Medium for Dynamically Deducing Electricity Sales Considering Domain Knowledge

Abstract

The present invention relates to a method, device and medium for dynamically deducing electricity sales volume considering domain knowledge. The method includes the following steps: collecting electricity sales volume data in the target domain to construct an adjacency matrix of industry relationships and analyzing the association strength between industries; collecting electricity sales volume data of upstream and downstream industries in the industrial chain of the target domain and analyzing the interaction relationship between upstream and downstream industries in the industrial chain; constructing a feature extraction model to extract enterprise maintenance features, policy adjustment features and market condition fluctuation features in the target domain; constructing a training set based on the above data and performing sample augmentation on the training set; constructing electricity sales boundary constraints based on grid power supply and equipment physical characteristics; constructing an electricity sales volume prediction model based on the electricity sales boundary constraints, training the electricity sales volume prediction model with the training set after sample augmentation, and predicting the electricity sales volume in the target domain through the trained electricity sales volume prediction model.

Description

Technical Field

[0001] The present invention relates to a method, device and medium for dynamically inferring electricity sales volume considering domain knowledge, and belongs to the technical field of electricity sales volume prediction. Background Technique

[0002] Electricity sales volume prediction is an important task in power system planning and operation. Common methods include traditional statistical methods, machine learning methods, deep learning methods, and hybrid model methods. Traditional statistical methods such as time series analysis (ARIMA, exponential smoothing method) and regression analysis method model historical data and are applicable to scenarios with stable data and obvious linear relationships. Machine learning methods such as support vector machine (SVM), decision tree (random forest, XGBoost), and artificial neural network (ANN) can process non-linear and multi-dimensional data and have strong adaptability, especially suitable for predicting large-scale complex data.

[0003] Deep learning methods such as long short-term memory network (LSTM), gated recurrent unit (GRU), and Transformer model can capture long-term dependencies and global features in time series data and are especially suitable for time series prediction of electricity sales volume.

[0004] In addition, the hybrid model method combines multiple methods. For example, first use a statistical model to predict the trend, and then use a machine learning or deep learning model to optimize the residual part to further improve the prediction accuracy.

[0005] External factors affecting electricity sales volume such as weather, economic level, population growth, and holiday effects are also often incorporated into the modeling consideration. For example, model by combining temperature, seasonal changes in electricity demand, GDP, and industrial structure data.

[0006] In recent years, the development of big data technology and cloud computing has also provided efficient data processing and computing support for electricity sales volume prediction; in practical applications, ARIMA, LSTM, or XGBoost is usually selected for short-term prediction, while regression analysis, economic models, or deep learning methods are mostly used for medium- and long-term prediction, and complex prediction tasks often combine multiple models for hybrid prediction to achieve higher accuracy.

[0007] Although the existing electricity sales prediction methods have achieved remarkable results, there are still some drawbacks and limitations. Traditional statistical methods such as ARIMA and regression analysis rely on the stationarity and linear assumptions of data, and it is difficult to handle complex non-linear relationships and sudden changes. Machine learning and deep learning methods can model non-linear problems, but they have high requirements for the quantity and quality of data, are vulnerable to noise data, have many model parameters, and the training process is complex, making it easy to overfit. In addition, deep learning models such as LSTM have a large amount of computation and low efficiency in long-term time series prediction, and the "black box" characteristics of the model make it difficult to interpret, affecting the credibility of practical applications. Although hybrid models can combine the advantages of multiple methods, the model design is complex, the parameter tuning is difficult, and the application cost is high. At the same time, the introduction of external factors such as weather, economic level, and policy changes can improve the prediction accuracy, but the data acquisition is difficult, and it is impossible to fully quantify and predict the impact of emergencies on electricity demand.

[0008] Therefore, how to improve the generalization ability, computational efficiency, and interpretability of the model and solve the problems of data incompleteness and uncertainty remains an important challenge in the field of electricity sales prediction. Summary of the Invention

[0009] In order to solve the problems existing in the above-mentioned prior art, the present invention proposes a dynamic deduction method, device, and medium for electricity sales considering domain knowledge.

[0010] The technical solution of the present invention is as follows:

[0011] On the one hand, the present invention provides a dynamic deduction method for electricity sales considering domain knowledge, including the following steps:

[0012] Collect the electricity sales data of the whole industry in multiple regions of the target domain, and construct an industry relationship adjacency matrix based on the electricity sales data to analyze the correlation intensity between industries;

[0013] Collect the electricity sales data of the upstream and downstream industries in the industrial chain of the target domain to construct an upstream and downstream relationship network model of the industrial chain, and analyze the interaction relationship between the upstream and downstream industries in the industrial chain;

[0014] Construct a feature extraction model to extract the enterprise maintenance characteristics, policy adjustment characteristics, and market condition fluctuation characteristics of the target domain;

[0015] Based on the electricity sales data of the whole industry in multiple regions of the target domain, the correlation intensity between industries, the interaction relationship between the upstream and downstream industries in the industrial chain, the enterprise maintenance characteristics, the policy adjustment characteristics, and the market condition fluctuation characteristics, construct a training set, and perform sample expansion on the training set;

[0016] Construct electricity sales boundary constraints based on grid power supply and equipment physical characteristics;

[0017] Construct a power sales prediction model based on power sales boundary constraints, train the power sales prediction model with the training set after sample augmentation, and predict the power sales in the target domain using the trained power sales prediction model.

[0018] As a preferred embodiment of the present invention, the steps for constructing the industry relationship adjacency matrix are as follows:

[0019] Construct a power sales feature matrix for different regions based on the power sales data of the entire industry in multiple regions of the target domain;

[0020] Define the dynamic correlation coefficient between regions based on the power sales feature matrix of different regions, as shown in the following formula:

[0021] ;

[0022] Where: represents the dynamic correlation coefficient between region and region ; represents the data sample in the power sales feature matrix of region at time ; represents the lag parameter; represents the data sample in the power sales feature matrix of region at time ; represents the standard deviation of the power sales feature matrix of region ;

[0023] Construct an industry association network based on the dynamic correlation coefficient between regions. The nodes in the industry association network represent an industry, and the edges represent the associations between industries. Then, construct an industry relationship adjacency matrix based on the industry association network, as shown in the following formula:

[0024] ;

[0025] Where: represents the connection strength between nodes and node in the industry relationship adjacency matrix, that is, the association strength between industry and industry ; represents the lag time interval; represents the start time; represents the dynamic correlation coefficient between the corresponding regions of nodes and node at time

[0026] As a preferred embodiment of the present invention, the specific steps for analyzing the functional relationship between the upstream and downstream industries of the industrial chain are as follows:

[0027] Construct a restricted model and an unrestricted model based on the electricity sales data of the upstream and downstream industries of the industrial chain in the target field, as shown in the following formula:

[0028] ;

[0029] ;

[0030] Where: represents the output result of the restricted model at time ; represents the number of lag order numbers of the electricity sales demand of the downstream industry; represents the -th lag order number of the electricity sales demand of the downstream industry; represents the electricity sales demand of the downstream industry at time represents the output result of the unrestricted model at time ; represents the number of lag order numbers of the electricity sales demand of the upstream industry; represents the -th lag order number of the electricity sales demand of the upstream industry; represents the electricity sales demand of the upstream industry at time represents the random error term; represents the -th lag coefficient of the electricity sales demand of the downstream industry; represents the -th lag coefficient of the electricity sales demand of the upstream industry;

[0031] Calculate the lag parameter value based on the residuals of the restricted model and the unrestricted model, as shown in the following formula:

[0032] ;

[0033] Where: represents the sum of squared residuals of the restricted model; represents the sum of squared residuals of the unrestricted model;

[0034] Construct an industrial chain upstream and downstream relationship network model based on a graph neural network, abstract each link in the upstream and downstream of the industrial chain as a node, and abstract the connection relationship between each link in the upstream and downstream of the industrial chain as an edge. The industrial chain upstream and downstream relationship network model is specifically shown in the following formula:

[0035] ;

[0036] Where: Represents the node features of the nth layer of the upstream and downstream relationship network model of the industrial chain, including the connection strength between each node; Represents the Laplacian matrix; Represents the adjacency matrix, which is constructed based on the industrial relationship adjacency matrix and obtained by substituting the calculated lag parameter; Represents the node features of the nth layer of the upstream and downstream relationship network model of the industrial chain; Represents the weight matrix of the nth layer of the upstream and downstream relationship network model of the industrial chain; Represents the sigmoid function;

[0037] Analyze the role relationship of the upstream and downstream industries of the industrial chain based on the final output of the upstream and downstream relationship network model of the industrial chain.

[0038] As a preferred implementation manner of the present invention, the specific extraction steps of the enterprise maintenance characteristics, policy adjustment characteristics, and market condition fluctuation characteristics are as follows:

[0039] Construct an enterprise maintenance feature extraction model based on a machine learning model;

[0040] Collect the prior knowledge of business personnel in the target field. The prior knowledge of business personnel in the target field specifically includes the enterprise operation data, equipment maintenance cycle, and equipment maintenance cycle in the target field. Based on the prior knowledge of business personnel in the target field, extract the enterprise maintenance characteristics through the enterprise maintenance feature extraction model, and represent them as a mask matrix:

[0041] ;

[0042] Where: Represents the enterprise maintenance feature mask matrix; Represents the start time; Represents the end time; Represents the indicator function; Represents any enterprise maintenance feature; Represents the nth Represents the enterprise maintenance feature value; Represents the enterprise maintenance feature value threshold;

[0043] Collect the policy adjustment data and market condition data in the target field, construct an impact quantification model based on a machine learning model, and extract the policy adjustment characteristics and market condition fluctuation characteristics in the target field through the impact quantification model, as shown in the following formula:

[0044] ;

[0045] ;

[0046] Wherein: represents the policy adjustment feature; represents the policy adjustment data sample at time; represents the influence intensity corresponding to the policy adjustment data sample; represents the attenuation coefficient corresponding to the policy adjustment data sample; represents the market condition fluctuation feature; represents the market condition data sample at time; represents the influence intensity corresponding to the market condition data sample; represents the attenuation coefficient corresponding to the market condition data sample.

[0047] As a preferred embodiment of the present invention, the specific steps for sample expansion of the training set are as follows:

[0048] Perform translation, scaling, and interpolation operations on the training set to obtain an enhanced training set;

[0049] Construct a historical event database, which contains various types of historical impact events, and calculate the impact of historical impact events on electricity sales, as shown in the following formula:

[0050] ;

[0051] Wherein: represents the impact of the historical impact event at time on electricity sales prediction; represents the electricity sales data of the entire industry in multiple regions of the target field at time; represents the impact coefficient of the th historical impact event on the electricity sales data of the entire industry in multiple regions of the target field; represents the influence function; represents the occurrence time of the th historical impact event;

[0052] Generate various types of historical impact events through simulation, calculate the impact of the historical impact events after simulation on electricity sales based on the above formula, and then weight each historical impact event;

[0053] Incorporate the impact of the historical impact events on electricity sales before and after generation into the enhanced training set to obtain a training set containing influencing factors;

[0054] Generate random Gaussian noise for the training set containing influencing factors based on noise perturbation to obtain a training set after noise generation;

[0055] Set different resampling frequencies according to the weight sizes of each historical impact event in the training set after noise generation, and resample the remaining data in the training set at the time nodes corresponding to the historical impact events according to the resampling frequencies to obtain an expanded training set with samples.

[0056] As a preferred embodiment of the present invention, a power sales boundary constraint is constructed based on grid power supply and equipment physical characteristics, as shown in the following formula:

[0057] ;

[0058] ;

[0059] ;

[0060] ;

[0061] Where: Represents the power generation capacity of the power grid At time Represents the power grid Transmission capacity at time Represents Power sales volume at time Represents the power transmission line capacity constraint of the power grid; Represents the average voltage of the power node ; Represents the power node Average voltage; Represents the power node And the power node Line reactance between; Represents the power node And the power node Voltage phase angle difference between.

[0062] As a preferred embodiment of the present invention, the following constraints are set on the basis of the power sales boundary constraint:

[0063] Line loss constraint:

[0064] ;

[0065] Where: Represents the power line loss; Represents the power line current; Represents the power line resistance; Represents the power line loss threshold;

[0066] Voltage stability constraint:

[0067] ;

[0068] Wherein: represents the minimum allowable voltage of the power grid; represents the maximum allowable voltage of the power grid; represents the real-time voltage of the power grid;

[0069] Transformer capacity constraint:

[0070] ;

[0071] Wherein: represents the transformer capacity; represents the average power transmission capacity of the power grid.

[0072] As a preferred embodiment of the present invention, taking the power sales boundary constraint, line loss constraint, voltage stability constraint, and transformer capacity constraint as the prior information of the power sales prediction model, the loss function of the power sales prediction model is:

[0073] ;

[0074] Wherein: represents the loss function of the data part of the power sales prediction model; represents the loss function of the power sales boundary constraint; represents the loss function of the voltage stability constraint; represents the loss function of the line loss constraint; represents the loss function of the transformer capacity constraint; , , , represent weight parameters.

[0075] On the other hand, the present invention also provides an electronic device, including a memory, a processor, and a computer program stored on the memory and executable on the processor. When the processor executes the program, the method described in any embodiment of the present invention is implemented.

[0076] On yet another aspect, the present invention also provides a computer-readable storage medium, on which a computer program is stored. When the program is executed by a processor, the method described in any embodiment of the present invention is implemented.

[0077] The present invention has the following beneficial effects:

[0078] By integrating multi-source data, physical constraints, and key event information, the present invention can comprehensively capture the complex factors affecting electricity sales volume, accurately identify the long-term trends, seasonality, and sudden fluctuations in the time series, significantly improve the accuracy of the prediction results and the stability under different scenarios, help power enterprises optimize resource allocation, identify potential risks of supply-demand imbalance in advance, enhance the scientificity and efficiency of power dispatching, and thus achieve the safe and stable operation of the power system and the maximization of economic benefits. BRIEF DESCRIPTION OF THE DRAWINGS

[0079] Figure 1 It is a flowchart of the method of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0080] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0081] It should be understood that the step numbers used in the text are only for convenient description and do not limit the execution order of the steps.

[0082] It should be understood that the terms used in the specification of the present invention are only for the purpose of describing specific embodiments and are not intended to limit the present invention. As used in the specification of the present invention and the appended claims, unless the context clearly indicates otherwise, the singular forms "a", "an", and "the" are intended to include the plural forms.

[0083] The terms "comprising" and "including" indicate the presence of the described features, wholes, steps, operations, elements, and / or components, but do not exclude the presence or addition of one or more other features, wholes, steps, operations, elements, components, and / or their combinations.

[0084] The term "and / or" refers to any combination and all possible combinations of one or more of the associated listed items, and includes these combinations.

[0085] Embodiment 1:

[0086] Refer to Figure 1 , a dynamic deduction method for electricity sales volume considering domain knowledge, including the following steps:

[0087] Collect the electricity sales volume data of the whole industry in multiple regions of the target domain, and construct an industry relationship adjacency matrix based on the electricity sales volume data to analyze the association strength between industries;

[0088] Collect the electricity sales data of the upstream and downstream industries in the industrial chain of the target field to construct a relationship network model of the upstream and downstream industries in the industrial chain, and analyze the interaction relationship between the upstream and downstream industries in the industrial chain;

[0089] Construct a feature extraction model to extract the enterprise maintenance features, policy adjustment features, and market condition fluctuation features of the target field;

[0090] Based on the electricity sales data of the entire industry in multiple regions of the target field, the association strength between industries, the interaction relationship between the upstream and downstream industries in the industrial chain, the enterprise maintenance features, policy adjustment features, and market condition fluctuation features, construct a training set, and perform sample expansion on the training set;

[0091] Construct electricity sales boundary constraints based on grid power supply and equipment physical characteristics;

[0092] Based on the electricity sales boundary constraints, construct an electricity sales prediction model, train the electricity sales prediction model through the training set after sample expansion, and predict the electricity sales of the target field through the trained electricity sales prediction model.

[0093] As a preferred implementation manner of this embodiment, the construction steps of the industry relationship adjacency matrix are as follows:

[0094] Based on the electricity sales data of the entire industry in multiple regions of the target field, construct electricity sales feature matrices for different regions;

[0095] Define the dynamic correlation coefficient between regions based on the electricity sales feature matrices of different regions, as shown in the following formula:

[0096] ;

[0097] Where: represents region and region the dynamic correlation coefficient between them; represents at time the data sample in the electricity sales feature matrix of region represents the lag parameter, which is used to analyze the conduction time lag of electricity sales; represents at time the data sample in the electricity sales feature matrix of region represents region the standard deviation of the electricity sales feature matrix; represents region the standard deviation of the electricity sales feature matrix;

[0098] Construct an industry association network based on the dynamic correlation coefficient between regions. The nodes in the industry association network represent an industry, and the edges represent the associations between industries. Then, construct an industry relationship adjacency matrix based on the industry association network, as shown in the following formula:

[0099] ;

[0100] Where: represents the connection strength between node and node in the industry relationship adjacency matrix, that is, the association strength between industry and industry ; represents the lag time interval; represents the start time; represents the dynamic correlation coefficient between the regions corresponding to node and node at time

[0101] Through multi-dimensional cross-comparison, reveal the interdependence relationship and potential laws of electricity sales volume among different regions and industries, clearly display the spatial distribution and evolution trend of the electricity sales volume association relationship among different regions and industries, so as to provide support for the optimal allocation of power resources.

[0102] As a preferred implementation manner of this embodiment, the specific steps for analyzing the role relationship of the upstream and downstream industries in the industrial chain are as follows:

[0103] Construct a restricted model (which means not considering the influence of upstream enterprises) and an unrestricted model (which means considering the influence of upstream enterprises) based on the electricity sales volume data of the upstream and downstream industries in the target field industrial chain, as shown in the following formula:

[0104] ;

[0105] ;

[0106] Where: represents the output result of the restricted model at time represents the number of lag order numbers of the electricity sales demand of the downstream industry; represents the th lag order number of the electricity sales demand of the downstream industry; represents the electricity sales demand of the downstream industry at time represents the output result of the unrestricted model at time represents the number of lag order numbers of the electricity sales demand of the upstream industry; represents the The number of lag orders; represents the electricity sales demand of the upstream industry at time represents the random error term; represents the th lag coefficient of the electricity sales demand of the downstream industry, indicating the lagging impact of the downstream electricity sales demand itself; represents the th lag coefficient of the electricity sales demand of the upstream industry, indicating the lagging impact of the upstream industry's electricity sales demand itself;

[0107] Calculate the lag parameter values based on the residuals of the restricted model and the unrestricted model, as shown in the following formula:

[0108] ;

[0109] where: represents the sum of squared residuals of the restricted model; represents the sum of squared residuals of the unrestricted model;

[0110] Construct an upstream and downstream industrial chain relationship network model based on a graph neural network, abstract each link in the upstream and downstream of the industrial chain as a node, and abstract the connection relationship between each link in the upstream and downstream of the industrial chain as an edge. The upstream and downstream industrial chain relationship network model is specifically shown in the following formula:

[0111] ;

[0112] where: represents the node features of the th layer of the upstream and downstream industrial chain relationship network model, including the connection strength between each node; represents the Laplacian matrix; represents the adjacency matrix, which is constructed based on the industry relationship adjacency matrix and obtained by substituting the calculated lag parameters; represents the node features of the th layer of the upstream and downstream industrial chain relationship network model; represents the weight matrix of the th layer of the upstream and downstream industrial chain relationship network model; represents the sigmoid function;

[0113] With the help of a graph neural network, upgrade the traditional input-output table to a dynamically learnable graph structure, further capture the potential dependency patterns in time series data, and automatically identify which changes in upstream links have the greatest impact on the downstream, as well as the duration and magnitude of the impact;

[0114] Analyze the interaction relationship between the upstream and downstream industries of the industrial chain based on the final output of the upstream and downstream industrial chain relationship network model.

[0115] As a preferred implementation of this embodiment, the specific extraction steps of the enterprise maintenance characteristics, policy adjustment characteristics, and market condition fluctuation characteristics are as follows:

[0116] Construct an enterprise maintenance feature extraction model based on a machine learning model;

[0117] Collect the prior knowledge of business personnel in the target field. The prior knowledge of business personnel in the target field specifically includes enterprise operation data, equipment maintenance cycle, and equipment overhaul cycle in the target field. Based on the prior knowledge of business personnel in the target field, extract enterprise maintenance characteristics through the enterprise maintenance feature extraction model, and represent them as a mask matrix:

[0118] ;

[0119] Where: Represents the enterprise maintenance feature mask matrix; Represents the start time; Represents the end time; Represents the indicator function; Represents any enterprise maintenance feature; Represents the th enterprise maintenance feature; Represents the enterprise maintenance feature value; Represents the enterprise maintenance feature value threshold;

[0120] Collect the policy adjustment data and market condition data (such as changes in commodity prices) in the target field. Construct an impact quantification model based on a machine learning model, and extract the policy adjustment characteristics and market condition fluctuation characteristics of the target field through the impact quantification model, as shown in the following formula:

[0121] ;

[0122] ;

[0123] Where: Represents the policy adjustment characteristic; Represents The policy adjustment data sample at time Represents the impact intensity corresponding to the policy adjustment data sample; Represents the attenuation coefficient corresponding to the policy adjustment data sample; Represents the market condition fluctuation characteristic; Represents The market condition data sample at time Represents the impact intensity corresponding to the market condition data sample; Represents the attenuation coefficient corresponding to the market condition data sample.

[0124] As a preferred implementation manner of this embodiment, the specific steps for sample augmentation of the training set are as follows:

[0125] Perform translation, scaling, and interpolation operations on the training set to obtain a data-augmented training set;

[0126] Construct a historical event database, which contains various types of historical impact events (such as equipment failures, maintenance plans, natural disasters, policy adjustments, and market abnormal fluctuations, etc.), and calculate the impact of historical impact events on electricity sales volume, as shown in the following formula:

[0127] ;

[0128] Where: represents the impact of the historical impact event at time on the electricity sales volume prediction; represents the electricity sales volume data of the entire industry in multiple regions of the target domain at time ; represents the influence coefficient of the th historical impact event on the electricity sales volume data of the entire industry in multiple regions of the target domain; represents the influence function; represents the occurrence time of the

[0129] th historical impact event;

[0130] Generate various types of historical impact events through simulation, calculate the impact of the historical impact events after simulation generation on electricity sales volume based on the above formula, and then assign weights to each historical impact event;

[0131] For some scarce but important events (such as extreme weather or large-scale power outages), use a rule-driven method to generate synthetic data to supplement the deficiencies of the training set;

[0132] Incorporate the impacts of the historical impact events before and after generation on electricity sales volume into the data-augmented training set to obtain a training set containing influencing factors;

[0133] Generate random Gaussian noise for the training set containing influencing factors based on noise perturbation to obtain a training set after noise generation;

[0134] According to the weight size of each historical impact event in the training set after noise generation, set different resampling frequencies, and resample the remaining data in the training set at the time nodes corresponding to the historical impact events according to the resampling frequencies to obtain a training set after sample augmentation.

[0135] As a preferred implementation manner of this embodiment, construct a power sales boundary constraint based on grid power supply and equipment physical characteristics, as shown in the following formula:

[0135] ;

[0136] ;

[0137] ;

[0138] ;

[0139] Wherein: represents the power generation capacity of the power grid at time ; represents the power transmission capacity of the power grid at time ; represents the electricity sales volume at time the capacity constraint of the power transmission line of the power grid; represents the average voltage of the power node ; represents the average voltage of the power node ; represents the power node and the power node the line reactance between; represents the power node and the power node the voltage phase angle difference between.

[0140] As a preferred embodiment of this embodiment, the following constraints are set on the basis of the electricity sales boundary constraint:

[0141] Line loss constraint:

[0142] ;

[0143] Wherein: represents the power line loss; represents the power line current; represents the power line resistance; represents the power line loss threshold;

[0144] Voltage stability constraint:

[0145] ;

[0146] Wherein: represents the minimum allowable voltage of the power grid; represents the maximum allowable voltage of the power grid; represents the real-time voltage of the power grid;

[0147] Transformer capacity constraint:

[0148] ;

[0149] Among them: represents the transformer capacity; represents the average power transmission capacity of the power grid.

[0150] As a preferred implementation manner of this embodiment, taking the power sales boundary constraint, line loss constraint, voltage stability constraint, and transformer capacity constraint as the prior information of the power sales prediction model, the loss function of the power sales prediction model is:

[0151] ;

[0152] Among them: represents the loss function of the data part of the power sales prediction model; represents the loss function of the power sales boundary constraint; represents the loss function of the voltage stability constraint; represents the loss function of the line loss constraint; represents the loss function of the transformer capacity constraint; , , , represent weight parameters;

[0153] The specific loss function of the data part is:

[0154] ;

[0155] Among them: represents the actual value of the power sales volume at time identifies the predicted value of the power sales volume of the power sales prediction model at time t.

[0156] Embodiment 2:

[0157] This embodiment provides an electronic device, including a memory, a processor, and a computer program stored on the memory and executable on the processor. When the processor executes the program, the method described in any embodiment of the present invention is implemented.

[0158] Embodiment 3:

[0159] This embodiment provides a computer-readable storage medium, on which a computer program is stored. When the program is executed by a processor, the method described in any embodiment of the present invention is implemented.

[0160] In the embodiments of the present application, "at least one" means one or more, and "a plurality" means two or more. "And / or" describes the association relationship of associated objects, indicating that three relationships can exist. For example, A and / or B can represent the cases of A existing alone, A and B existing simultaneously, and B existing alone. Where A and B can be singular or plural. The character " / " generally represents an "or" relationship between the associated objects before and after. "At least one of the following" and its similar expressions refer to any combination of these items, including any combination of single or plural items. For example, at least one of a, b, and c can represent: a, b, c, a and b, a and c, b and c, or a and b and c, where a, b, and c can be single or multiple.

[0161] Those of ordinary skill in the art can realize that the various units and algorithm steps described in the embodiments disclosed herein can be implemented by a combination of electronic hardware, computer software, and electronic hardware. Whether these functions are executed in a hardware or software manner depends on the specific application and design constraints of the technical solution. Professional technicians can use different methods for each specific application to implement the described functions, but such implementation should not be considered to exceed the scope of the present application.

[0162] Those skilled in the art can clearly understand that for the convenience and simplicity of description, the specific working processes of the systems, devices, and units described above can refer to the corresponding processes in the foregoing method embodiments and will not be elaborated herein.

[0163] In several embodiments provided by the present application, if any function is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on such an understanding, the technical solution of the present application, in essence, or the part that contributes to the prior art, or a part of this technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to enable a computer device (which can be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present application. The foregoing storage medium includes: various media such as USB flash drives, mobile hard disks, read-only memories (hereinafter referred to as ROM), random access memories (hereinafter referred to as RAM), magnetic disks, or optical discs that can store program codes.

[0164] The above are only the embodiments of the present invention and do not limit the patent scope of the present invention. Any equivalent structure or equivalent process transformation made by using the specification and drawings of the present invention, or directly or indirectly applied in other related technical fields, shall be equally included in the patent protection scope of the present invention.

Claims

1. A method for dynamic deduction of electricity sales volume considering domain knowledge, characterized in that: The following steps are involved: Collect electricity sales data from all industries in multiple regions in the target field, and build an industry relationship adjacency matrix based on the electricity sales data to analyze the correlation strength between industries; Collect electricity sales data of upstream and downstream industries in the target field industrial chain, build an upstream and downstream relationship network model of the industrial chain, and analyze the role relationship of upstream and downstream industries in the industrial chain; the specific steps of analyzing the role relationship of upstream and downstream industries in the industrial chain are: Based on the electricity sales data of upstream and downstream industries in the target field industrial chain, a restricted model and an unrestricted model are constructed, as shown in the following formula: ； ； in: Representing a restricted model Output results at the moment; Indicates the number of lag orders of electricity sales demand in downstream industries; The first The number of lag orders; express The electricity sales demand of downstream industries at the moment; Represents an unrestricted model Output results at the moment; Indicates the number of lag orders of electricity sales demand in the upstream industry; The first The number of lag orders; express The electricity sales demand of upstream industries at all times; represents the random error term; The first hysteresis coefficient; The first hysteresis coefficient; The lagged parameter values ​​are calculated based on the residuals of the restricted model and the unrestricted model, as shown in the following formula: ； in: represents the residual sum of squares of the restricted model; represents the residual sum of squares of the unrestricted model; represents the lag time interval; Based on the graph neural network, a network model of upstream and downstream relationships in the industrial chain is constructed. Each link in the upstream and downstream of the industrial chain is abstracted as a node, and the connection relationship between each link in the upstream and downstream of the industrial chain is abstracted as an edge. The upstream and downstream relationship network model of the industrial chain is specifically shown in the following formula: ； in: The network model of the upstream and downstream relationship of the industrial chain The node characteristics of the layer, including the connection strength between each node; represents the Laplacian matrix; represents the adjacency matrix, which is constructed based on the industry relationship adjacency matrix and obtained by substituting the calculated lag parameters; The network model of the upstream and downstream relationship of the industrial chain Node characteristics of the layer; The network model of the upstream and downstream relationship of the industrial chain The weight matrix of the layer; Represents the sigmoid function; Analyze the relationship between upstream and downstream industries in the industrial chain based on the final output of the upstream and downstream relationship network model of the industrial chain; Construct a feature extraction model to extract enterprise maintenance features, policy adjustment features, and market fluctuation features in the target field; The training set is constructed based on the electricity sales data of all industries in multiple regions of the target field, the correlation strength between industries, the relationship between upstream and downstream industries in the industrial chain, the maintenance characteristics of enterprises, the characteristics of policy adjustments, and the characteristics of market fluctuations, and the samples of the training set are expanded; Establish power sales boundary constraints based on power grid supply and physical characteristics of equipment; A power sales prediction model is constructed based on the power sales boundary constraints, and the power sales prediction model is trained using the training set after sample expansion. The trained power sales prediction model is used to predict the power sales in the target area.

2. According to claim 1, a method for dynamically deducing power sales volume considering domain knowledge is characterized in that: The steps for constructing the industry relationship adjacency matrix are: Based on the electricity sales data of all industries in multiple regions in the target field, a characteristic matrix of electricity sales in different regions is constructed; Based on the characteristic matrix of electricity sales in different regions, the dynamic correlation coefficient between regions is defined as follows: ； in: Indicates the region and region The dynamic correlation coefficient between express Time Region Data samples in the electricity sales feature matrix; represents the hysteresis parameter; express Time Region Data samples in the electricity sales feature matrix; Indicates the region The standard deviation of the electricity sales characteristic matrix; Indicates the region The standard deviation of the electricity sales characteristic matrix; An industry association network is constructed based on the dynamic correlation coefficient between regions. The nodes in the industry association network represent an industry, and the edges represent the association between industries. Then, an industry relationship adjacency matrix is ​​constructed based on the industry association network, as shown in the following formula: ； in: Represents the nodes in the industry relationship adjacency matrix With Node The connection strength of the industry With the industry The strength of association; Indicates the start time; express Time Node With Node The dynamic correlation coefficient between corresponding regions.

3. According to claim 1, a method for dynamically deducing power sales volume considering domain knowledge is characterized in that: The specific steps for extracting the enterprise maintenance characteristics, policy adjustment characteristics and market fluctuation characteristics are as follows: Build an enterprise maintenance feature extraction model based on machine learning model; The prior knowledge of the business personnel in the target field is collected. The prior knowledge of the business personnel in the target field specifically includes the enterprise operation data, equipment maintenance cycle, and equipment overhaul cycle in the target field. Based on the prior knowledge of the business personnel in the target field, the enterprise overhaul features are extracted through the enterprise overhaul feature extraction model, which is expressed as a mask matrix: ； in: represents the enterprise maintenance feature mask matrix; Indicates the start time; Indicates the end time; represents the indicator function; Indicates any enterprise maintenance characteristics; Indicates Maintenance characteristics of each enterprise; represents the maintenance characteristic value of the enterprise; Indicates the threshold value of enterprise maintenance characteristic value; Collect policy adjustment data and market data in the target area, build an impact quantification model based on the machine learning model, and extract the policy adjustment characteristics and market fluctuation characteristics of the target area through the impact quantification model, as shown in the following formula: ； ； in: Indicates the characteristics of policy adjustments; express Policy adjustment data samples at each moment; It indicates the impact intensity corresponding to the policy adjustment data sample; represents the attenuation coefficient corresponding to the policy adjustment data sample; Indicates the characteristics of market fluctuations; express Market data samples at the moment; Indicates the impact intensity corresponding to the market data sample; Represents the attenuation coefficient corresponding to the market data sample.

4. According to claim 1, a method for dynamically deducing power sales volume considering domain knowledge is characterized in that: The specific steps of expanding the sample of the training set are: Perform translation, scaling, and interpolation operations on the training set to obtain a data-enhanced training set; A historical event database is constructed, which contains various types of historical impact events, and the impact of historical impact events on electricity sales is calculated, as shown in the following formula: ； in: express The impact of historical events on electricity sales forecast; express The target area is the electricity sales data of multiple regions and industries at all times; Indicates The impact coefficient of the historical impact event on the electricity sales data of multiple regions and industries in the target field; represents the influence function; Indicates The moment when a historically influential event occurred; Generate various types of historical impact events through simulation, and calculate the impact of the historical impact events generated by simulation on the electricity sales based on the above formula, and then assign weights to each historical impact event; The impact of historical events before and after the generation on electricity sales is included in the data-enhanced training set to obtain a training set containing influencing factors; Based on the noise perturbation, random Gaussian noise is generated for the training set containing the influencing factors to obtain the training set after noise generation; Different resampling frequencies are set according to the weight corresponding to each historical impact event in the training set after noise generation, and the remaining data in the training set are resampled at the time nodes corresponding to the historical impact events according to the resampling frequencies to obtain the training set after sample expansion.

5. According to claim 1, a method for dynamically deducing power sales volume considering domain knowledge is characterized in that: The boundary constraints for electricity sales are constructed based on the power supply of the power grid and the physical characteristics of the equipment, as shown in the following formula: ； ； ； ； in: Indicates the power grid The power generation capacity at any time; Indicates the power grid The power transmission capacity at any time; express The amount of electricity sold at the time; represents the power transmission line capacity constraints of the power grid; Represents a power node The average voltage of Represents a power node The average voltage of Represents a power node With power node The line reactance between Represents a power node With power node The voltage phase angle difference between them.

6. A method for dynamically deducing power sales volume considering domain knowledge according to claim 5, characterized in that: The following constraints are set based on the power sales boundary constraints: Line loss constraints: ； in: Indicates power line loss; Indicates the power line current; Indicates the resistance of the power line; Indicates the power line loss threshold; Voltage stability constraints: ； in: Indicates the minimum allowable voltage of the power grid; Indicates the maximum allowable voltage of the power grid; Indicates the real-time voltage of the power grid; Transformer capacity constraints: ； in: Indicates transformer capacity; Represents the average transmission capacity of the power grid.

7. The method for dynamically deducing power sales volume considering domain knowledge according to claim 6 is characterized in that: Taking the power sales boundary constraint, line loss constraint, voltage stability constraint and transformer capacity constraint as the prior information of the power sales forecasting model, the loss function of the power sales forecasting model is for: ； in: represents the loss function of the data part of the electricity sales prediction model; represents the boundary constraint loss function of electricity sales; represents the voltage stability constraint loss function; represents the line loss constraint loss function; represents the transformer capacity constraint loss function; , , , Represents the weight parameter.

8. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the program, the method according to any one of claims 1 to 7 is implemented.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the program is executed by a processor, the method according to any one of claims 1 to 7 is implemented.

Images